The Evolution of Bitcoin Mining to AI

In 2022, we were building a data center project in Florida. Everything was going great. Who doesn't like a project in Florida, right? But then in late September, my phone started blowing up. It was our customer calling us. Hurricane Ian hit mainland.

​

Who remembers? Because I sure do. We're talking Category 5, 160-mile-per-hour winds, and about $112 billion in damage.

​

The next morning, speaking to our customer, it was chaos. There were people everywhere on boats and Sea-Doos trying to help their neighbors. We all understood it was a crisis, and the priority was making sure everybody was safe and in good care. But here's the thing: the world doesn't stop. It doesn't stop for much, usually, even in these crises. There are still projects to complete. The show must go on. That's what AI demands.

​

Can your field construction and operation keep going in a hurricane? If your construction is halfway through, what survives? How many months of delays and repairs? If you're operational, can it keep running? Is it rated for the area where you deploy? Do you have redundancy and concurrently maintainable systems to meet your customer's expectations?

​

When I say hurricane, I don't just mean a literal hurricane. I'm talking about rain, mud, dust, and extreme hot and cold weather patterns that we're all facing on construction sites. What about supply chain interruptions and water supply interruptions? I've seen projects crippled for less than that.

​

Here's the thing. During the hurricane, we never stopped. We kept going. By the time they were doing cleanup, we were deploying phase two of the project.

​

My company has built infrastructure since 2011. We've been part of the Bitcoin industry since 2019, and I've gotten to see firsthand the group of high achievers that it's made of. You find land, you find power, and you put compute online faster than anyone. We've built lots of projects together, which I'm very proud of.

​

But AI demands more. We're not just serving the blockchain anymore. The good enough mentality and basic tier-one infrastructure are simply incompatible with an AI transition. We'll need to follow strict guidelines and operating processes to make these projects a success.

​

As we pivot and diversify to AI, let's consider four main areas of challenge: speed, scale, flexibility, and quality.

​

Traditional data centers are slow. We're talking 24 to 36 months of infrastructure build time, without considering bad weather, supply chain issues, labor shortages, or complex permitting. AI moves too fast for this. If we follow NVIDIA's GPU roadmap, your infrastructure will be a dinosaur, completely outdated by the time it's constructed.

​

Stick-built data center projects tie up large amounts of capital upfront. They're inflexible. It's large building infrastructure with fixed, centralized power and cooling systems. This is all based on the available GPU at the time of design, which may have been three years ago. It doesn't allow you to grow in phases as customer demand comes in.

​

Traditional data centers are inflexible by nature. You think you can retrofit for new generations of GPUs as they get released, but you have to think twice. Bitcoin miner upgrades follow the halving every four years. In between, you have minimal hash rate upgrades and very minimal changes on the hardware. You don't have to retrofit your system as you go.

​

But AI GPUs evolve every 12 months. We're not talking simple optimization here. We're talking about doubling the compute density, potentially changing the form factor, and shifting voltage as well. Those are major changes. At this rate, you would be in a constant retrofit pattern. You need to get ahead of the hardware cycle. You need infrastructure that allows flexibility for AI densification.

​

We all know that infrastructure quality in the Bitcoin mining industry varies a lot. Maybe a little too much. A lot of the time, it's due to CapEx constraints or the good enough mentality out there. But now we have a whole new set of rules.

​

AI customers will have different but very strict expectations. We're talking quick deployment, no single point of failure, concurrently maintainable systems, and the highest-quality components. We'll have to follow NVIDIA reference guidelines on top of that, which is not easy. This is a completely different ballgame.

​

Here's the response to our challenges: modular. A fully integrated modular system that includes everything from utility connection through your white space, cooling plant, power generation, power transformation, control room, warehousing, and a true modular AI factory built in a controlled factory environment.

​

Let's bring this back to the hurricane story for a second. The only way we maintained operation during the hurricane and delivered a project on time was a high-quality, pre-engineered modular system, engineered and built for the environment it was deployed in. In that case, hurricane rated.

​

Moving forward, let's think hurricane winds, high elevation operation, and seismic rating. A manufacturing facility with lean processes and quality processes will always win the race. Speed, scale, flexibility, and quality. Collectively, let's start thinking modular. That's how you control all the variables.

​

Let's forget 24 to 36 months. Keep this number in mind: 12 months and under. This is how you're going to attract customers, and this is how you're going to make your project successful. Modular infrastructure transforms data center delivery from a construction project to a product line with a predictable supply chain and manufacturing schedule. It allows a parallel path between site construction and manufacturing activity. That's how you're moving fast. The only path to AI-ready infrastructure at the speed the market demands is modular. Field construction simply can't keep pace.

​

Grow in phases. Scale as you grow. No guesswork, no overbuilding, no stranded capital in large building infrastructure. You start generating returns under 12 months. The scalability of a modular system is simply unmatched. You scale block by block without disrupting uptime and the operation of your previous phases.

​

Keep your modular design flexible. You can't allow hardware densification to set you back with retrofits all the time. We're talking adaptable power and cooling systems that can follow NVIDIA's GPU roadmap. Power draws are leaping from 130 to 600 kW plus. Voltage is shifting from 415 AC to 800 volt DC. Inflexible infrastructure doesn't age; it literally expires. Modular flexibility is not an option anymore. It's how you avoid building tomorrow's outdated data centers.

​

Building in a controlled factory environment allows for quality control and factory acceptance testing. You can follow all of these processes and execute them consistently, again mitigating surprises on site that could cost you on quality and overall project delivery.

​

Quality comes at a price. Typical Bitcoin mining infrastructure ranges between $150,000 and $500,000 per megawatt. An AI factory starts around $7.5 million per megawatt. This is the cost of an AI transition. I don't want to sound redundant, but it is a completely different ballgame. This is not the same industry, but there are ways to transition properly.

​

Let's go through the digital twin of our Octopod solution as an example of a modular solution. Octopod from Intelliflex is a true start-to-finish modular AI factory.

​

The first piece is the connecting module, where you bring all your medium-voltage power distribution system through busway. This is where you house all your fire suppression systems and control panels. This is also how you interconnect the next block when you grow your facilities.

​

Then we go through a contained hot aisle in the IT hall. This is where we have our cooling units and fan wall. They take care of all the residual heat that you have in your direct liquid cooling racks, and they take care of all your lower-density networking racks and storage that don't have liquid cooling.

​

In the cold aisle, we have all of our facility water coming through. This is our distributed pump system. You're going to have a one-to-one ratio with the IT rack, your GPUs, and the cooling equipment. This is all N+1 redundancy. As you can see, this is a facility. This is not a containerized component. We're talking about a true AI factory that you can walk through, install, operate for your day-to-day operations, and access every piece of equipment.

​

Modular always brings repeatability. The second data hall mirrors the other side. Overall, in one of our factories, we're going to have up to four of those, which will give you 8.8 MW of critical IT load.

​

The cooling plant has four levels. The plant is made with chiller units on the first floor in a six-to-make-five configuration, and it gives you all the chilled water cooling necessary to cool your IT loads. In the middle, we have our utility modules and pump systems. The secondary level has other utilities like power distribution.

​

Again, you'll see the repeatable modules on the other side. It's a mirror image. You have the three other chillers on this side. Outside, you can see the overall cooling plant from the outside, including levels three and four. It is very large, industrial, modular, and not containerized. It is all structural steel rated for the environment it's deployed in.

​

The dry coolers on top work in parallel with the chiller units, and they have the same configuration, a six-to-make-five configuration.

​

You need access to all your components. This is why we have walkways on the second floor. This is where we host all of our HVAC plenum, busway systems, tap boxes, and cable trays for networking. You can see all the tap boxes here for the direct liquid cooling racks, and four levels of cable tray for networking and fiber cable.

​

Again, you have access to every piece of equipment. We're talking about platforms to do all your preventive maintenance daily.

​

One last piece is our Mega Box. This is where you do the power transformation from medium voltage to low voltage distribution, which feeds the data center from there.

​

That's a quick overview of a modular data center, the Octopod from Intelliflex. It's an idea of how you can actually achieve this properly.

​

It doesn't matter how many times we go through it, it's always nice to see the digital twin. It really puts it in perspective.

​

Here's what I want everyone to remember: speed under 12 months. Scale block by block. Don't strand your capital. Flexibility means staying ahead of the hardware curve. Quality means building in a controlled manufacturing environment.

​

If you follow these four words, they will guide you toward a successful AI transition or buildout. Remember, modular wins the race every time: speed, scale, flexibility, and quality.

​

Thank you so much for listening. I'm truly looking forward to seeing you all on the floor. We're in booth 334, right across over here. We're displaying a lot of solutions, and you can take the digital twin for a spin. Thank you, and have a great rest of the conference.